System identification using the neural-extended Kalman filter for state-estimation and controller modification

Abstract

The neural extended Kalman filter (NEKF) is an adaptive state estimation technique that can be used in target tracking and directly in a feedback loop. It improves state estimates by learning the difference between the a priori model and the actual system dynamics. The neural network training occurs while the system is in operation. Often, however, due to stability concerns, such an adaptive component in the feedback loop is not considered desirable by the designer of a control system. Instead, the tuning of parameters is considered to be more acceptable. The ability of the NEKF to learn dynamics in an open-loop implementation, such as with target tracking and intercept prediction, can be used to identify mismodeled dynamics external to the closed-loop system. The improved nonlinear system model can then be used at given intervals to adapt the state estimator and the state feedback gains in the control law, providing better performance based on the actual system dynamics. This variation to neural extended Kalman filter control operations is introduced in this paper using applications to the nonlinear version of the standard cart-pendulum system.